Current ai implementations is that with (extraordinarily tedious training) they

Current ai implementations is that with (extraordinarily tedious training) they correctly categorize and prediction, and can then hand off to programmatic logic to ‘reason’ further.

Something like Go or Chess or Checkers, or even ‘training for image recognition’ is really not very complicated because the ‘world model’ for these programs is rather trivial, even ‘wayfinding’ through them is rather impressive.

The GPT-3 algorithm is interesting because it is so vast, and so accurately maintains associative context. That context’s world view is nonsensical however. It’s just word-patterns.

The hard problem – whch is where consciousness comes from – is constructing a world model, and predicting from that world model. Because humans (and anything that must reason like us) must have a ‘chessboard, go-board etc’ of the world, and the world is infinitely more cimplicated that those boards.

Worse, whatever we model is open to continuous revision.